Pith. sign in

REVIEW 4 major objections 5 minor 40 references

A Contrastive Diffusion-based Network (CDNet) for Time Series Classification

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read CDNet wraps a deep time series classifier with a CNN-based diffusion process that generates contrastive pairs, and claims it significantly improves accuracy on noisy, similar, and multimodal UCR binary datasets.

desk verdict A genuinely new idea for generating contrastive pairs via instance-to-instance diffusion, but the theoretical lemma is invalid and the empirical tables are internally inconsistent, so the paper needs major rework. read the letter →

arxiv 2507.21357 v1 pith:2JMXWUP7 submitted 2025-07-28 cs.LG

Yaoyu Zhang , Chi-Guhn Lee This is my paper
classification cs.LG
keywords timeseriesclassificationdiffusionmodelscontrastivelearningCNNreverseUCRArchiveuncertainty-weightedlossmultimodalitynoiserobustness
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

CDNet is a pre-training wrapper for deep time series classifiers. Its central claim is that a classifier's accuracy on hard binary univariate problems—where classes look alike, within-class shapes vary, and labels are noisy—can be lifted by first training the classifier with contrastive sample pairs generated by a diffusion process. The diffusion process is unusual: instead of denoising an individual sample back to itself, it interpolates between two real samples of the same class (positive pairs) and between samples of different classes (negative pairs), and a set of 1D CNNs learns the reverse transition. The paper reports that wrapping InceptionTime, 1DCNN, and LSTM_FCN this way improves their accuracy on the UCR Archive and that the gains grow as noise, inter-class similarity, and multimodality increase. If the empirical claim holds, CDNet would give a simple plug-in way to make existing deep TSC models more reliable in exactly the conditions where they currently struggle.

What carries the argument

The central object is the CNN-approximated reverse diffusion transition between time series instances. Starting from a forward process $x_t = \sqrt{1-\beta_t}\,x_{t-1} + (1-\sqrt{1-\beta_t})\,x^0_j + \sqrt{\beta_t}\,\epsilon$, CDNet trains a 1D CNN $f^t_{\theta_t}$ to map the noisy interpolated state $x_t$ back to $x_{t-1}$, once for within-class pairs and once for across-class pairs, with separate CNNs per class and per diffusion step. These trained reverse steps are then composed to generate positive samples $x^{t+}_i$ and negative samples $x^{t-}_i$ for contrastive losses. The uncertainty-weighted loss $L_{\text{total}} = \frac{1}{2\sigma^2_{\text{CE}}}L_{\text{CE}} + \frac{1}{2\sigma^2_{\text{SNN}}}L_{\text{SNN}} + \frac{1}{2\sigma^2_{\text{Triplet}}}L_{\text{Triplet}} + \log\sigma_{\text{CE}} + \log\sigma_{\text{SNN}} + \log\sigma_{\text{Triplet}}$ is what balances classification and contrastive objectives during pre-training.

What would settle it

Run CDNet's forward process with ten different random interpolation samples and noise draws for the same $x_t$, and check whether the resulting $x_{t-1}$ targets are nearly identical; if they scatter widely, no single CNN can approximate the reverse map with arbitrarily low error, and the theoretical grounding would need to be replaced. On the empirical side, an ablation that swaps the learned reverse CNNs for direct sample interpolation while keeping the loss unchanged would show whether the diffusion mechanism, rather than the contrastive pairs themselves, drives the reported accuracy gains.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the reverse process of a diffusion model can be repurposed as a contrastive data generator for time series classification. For each anchor sample, CDNet runs a forward process that linearly combines it with another sample—within the same class for positives, across classes for negatives—while adding Gaussian noise, then learns four stacks of 1D CNNs to reverse that process step by step. The authors claim this gives a theoretically grounded way to generate new samples that explore intra-class modes and expose inter-class confusion, and they support it with lemmas stating that CNN-based reverse diffusion recovers multimodal support and that 1D CNNs can approximate reverse transitions arbitrarily well. Trained with an uncertainty-weighted composite of cross-entropy, soft nearest-neighbor, and triplet losses, then fine-tuned on the last layer, CDNet is claimed to significantly improve InceptionTime, 1DCNN, and LSTM_FCN on the UCR binary datasets and to be the top-ranked method in critical-difference comparisons.

Load-bearing premise

The central theoretical guarantee assumes the reverse diffusion transition is a deterministic Lipschitz function of the noisy signal $x_t$, so a CNN can approximate it; in the actual forward process the target $x_{t-1}$ also depends on the randomly chosen interpolation sample and the noise draw, which are not determined by $x_t$.

Editorial extensions

If this is right

  • Wrapping a convolutional or recurrent deep classifier with CDNet pre-training can raise its accuracy on binary univariate time series without changing the classifier's architecture.
  • The performance gap between deep and non-deep TSC methods on ambiguous binary datasets can be narrowed by CDNet's generated contrastive pairs.
  • CDNet's advantage is largest when data are noisy, classes are similar, and within-class variation is multimodal, so it targets the conditions that most hurt deep classifiers.
  • The composable reverse CNNs generate trajectories rather than single augmented samples, which is what lets triplet and soft nearest-neighbor losses see a range of intra-class and inter-class variations.
  • Fine-tuning only the last layer after pre-training is sufficient to realize the gains, keeping the wrapper cheap to deploy.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • We infer that the same within-class and across-class interpolation idea could be applied to multivariate and multiclass time series by training more reverse chains and sharing parameters across classes, though the paper does not attempt this.
  • We infer that CDNet's benefit depends on the base classifier being convolutional; the paper's own results show smaller and sometimes negative changes for LSTM_FCN, which suggests the generated trajectories help most when the backbone already operates on local temporal differences.
  • We infer that the generated trajectories could serve as a generic augmentation for other time series tasks such as forecasting or anomaly detection, where hard positive and negative pairs are also useful, though this is untested.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes CDNet, a plug-in pre-training framework for univariate binary time series classification. CDNet augments a base deep classifier by generating within-class and across-class sample transitions through a forward diffusion process and approximating the reverse steps with 1D CNNs; the generated positive and negative trajectories are then used to train an uncertainty-weighted composite loss (cross-entropy, soft nearest neighbor, and triplet). The authors claim theoretical guarantees that CNNs approximate reverse diffusion transitions and support the multimodal class distribution, and they report significant accuracy gains over InceptionTime, 1DCNN, and LSTM_FCN on the UCR Archive and simulated datasets under noise, class similarity, and multimodality.

Significance. If the empirical results were reliable, CDNet would be a practically useful wrapper for improving strong deep TSC baselines with a modest change to the training pipeline, and the experiment design (all binary UCR datasets, simulated ablations, multiple backbones) is well targeted. The uncertainty-weighted loss and the idea of learning contrastive trajectories via diffusion are reasonable starting points. However, the theoretical foundation is invalid as stated, and the reported tables are internally inconsistent. Since the paper explicitly advertises 'theoretically grounded' reverse diffusion and 'consistent improvement' over SOTA models as its key contributions, the current manuscript does not establish either claim.

major comments (4)
  1. [Appendix A.1, Lemma 2 (and Section 3.2)] The reverse mapping is not a deterministic function of x_t. The equation defining the forward process, x_t = sqrt(1-beta_t) x_{t-1} + (1 - sqrt(1-beta_t)) x0_j + sqrt(beta_t) epsilon_t, cannot be solved for x_{t-1} as a function of x_t alone because x0_j and epsilon_t are independent random draws that are not determined by x_t. The proof then writes x_{t-1} = g(x_t) with g affine in x_t, but for a fixed x_t there are infinitely many possible x_{t-1} values indexed by (x0_j, epsilon_t). Consequently the statement sup_{x_t in F_M} ||f_theta(x_t) - x_{t-1}||_2 <= epsilon is not well-posed, and the application of the CNN universal approximation theorem has no valid target function. This invalidates the 'feasibility of CNNs' argument in Section 3.2 and the related claim in the contributions list.
  2. [Appendix A.1, Lemma 1] The proof of Lemma 1 assumes the conclusion it is meant to establish. It asserts that 'at convergence, f_theta maps each x_t toward its corresponding x, which comes from a particular mode P_k' and then concludes that the pushforward covers all mode supports. The step from convergence of SGD to a stationary point of a nonconvex objective to the claim that different neighborhoods of x_t are mapped to samples from every mode is not justified; convexity of ||.||_2^2 in f_theta alone does not imply the global minimizer has the stated support-coverage property. The lemma may be plausible, but the proof as written is circular and does not provide the advertised guarantee.
  3. [Appendix A.1, Lemmas 3 and 4] Both lemmas assume that eta, the noise component of x_{t-1} = s_{t-1} + eta, is independent of x_t. But x_t is defined through x_{t-1}, so eta is a component of x_t and the independence assumption is false. The derivation of the MMSE identity f_t(x_t) = E[x_{t-1} | x_t] and the subsequent decomposition with Var(eta) therefore do not go through as written. These lemmas also cannot repair Lemma 2, because the conditional expectation is a different object from the claimed deterministic function g(x_t).
  4. [Tables 1 and 2] The same CDNet-enhanced models report different accuracies for the same datasets across the two tables. For example, ECG200: CDNet_IT is 0.910 in Table 1 but 0.935 in Table 2; FordB: CDNet_IT is 0.891 in Table 1 but 0.927 in Table 2; Ham: CDNet_LSTM is 0.840 in Table 1 but 0.781 in Table 2. Additionally, several 'Change in Accuracy' columns are arithmetically inconsistent with the base and CDNet columns in Table 1; for instance, the BC row reports delta_LSTM = 0.194 while 0.994 - 0.550 = 0.444. Since the paper's central empirical claim is that CDNet significantly improves the baseline models, these internal inconsistencies make the reported results unreliable as evidence for that claim.
minor comments (5)
  1. [Algorithm 1 and Section 3.2] The notation for the second sample is inconsistent: the main text defines the forward process using x0_j, while Algorithm 1 line 6 writes x0_i; please align the indices throughout.
  2. [Section 3.3] The triplet loss definition contains a typo, 'Σ_{T i=i}', which should read 'Σ_{t=1}^{T}'.
  3. [Section 3.3] The SNN loss formula is difficult to parse: the '+epsilon' inside the logarithm and the definition of mask(i, j) are not clearly specified; please provide a clean formulation.
  4. [Section 4] No code, hyperparameter settings, or detailed training configuration for CDNet are provided, which limits reproducibility of the reported results.
  5. [Section 4.1.2 and Conclusion] The experimental comparison includes non-deep SOTA classifiers and loss-function variants, but not other contrastive or diffusion-based augmentation methods discussed in Section 2, such as TS2Vec or TS-TCC; the appended limitation statement acknowledges this, but the 'significant improvement' claim would be better supported by such comparisons.

Circularity Check

2 steps flagged · score 6.0 of 10

Partial circularity: Lemma 1's mode-coverage guarantee assumes the mode-recovery it claims to derive (and inherits its conclusion from the training-target support), while Lemma 2's reverse mapping g(xt) is constructed from the forward equation with random draws x0_j and εt still present, making the CNN-approximation claim an artifact of definition.

  1. self definitional [Appendix A.1, proof of Lemma 1 (lemma stated in Section 3.2, 'The feasibility of CNNs')]
    "At convergence, fθ maps each xt toward its corresponding x, which comes from a particular mode P(y)_k. Since qt(xt | y) includes interpolated/noisy mixtures between modes, fθ learns to map different neighborhoods in xt space to different x from different modes. Therefore, the pushforward distribution P̂θ(x | y) induced by sampling xt ∼ qt(xt | y) and applying fθ satisfies: supp(P̂θ(x | y)) ⊇ ∪_{k=1}^K supp(P(y)_k)."

    The proof's load-bearing step asserts the mode-recovery property the lemma must establish: 'fθ maps each xt toward its corresponding x, which comes from a particular mode' is exactly the claim that the reverse map sends each noisy neighborhood back to its own mode, and the support-inclusion conclusion is then read off from that assertion. Independently, the conclusion is forced by construction: the reconstruction objective minimizes E[||fθ(xt) − x||²] with x sampled from P(x|y), whose K modes all have π_k > 0, so any successful reconstruction has outputs ranging over ∪ supp(P_k) regardless of the CNN mechanism. The mode-coverage 'prediction' is therefore inherited from the training-target support and the assumed convergence, not derived from the architecture.

  2. self definitional [Appendix A.1, proof of Lemma 2 (forward process defined in Section 3.2)]
    "Solving for xt−1 yields the reverse mapping xt−1 = 1/√(1−βt) (xt − (1−√(1−βt)) x0_j − √(βt) εt), which we denote by g(xt) = xt−1. This function is affine in xt, and therefore Lipschitz continuous."

    The 'reverse mapping' g is manufactured by algebraically rearranging the forward equation, but the right-hand side still contains the random draws x0_j and εt, which are not determined by xt; for a fixed xt there are many possible xt−1 values, so no function g: R^M → R^M satisfies the displayed equation. The claimed Lipschitz constant 1/√(1−βt) and the sup-norm approximation bound sup_{xt∈FM} ||fθ(xt) − xt−1||₂ ≤ ε are therefore properties of a constructed object, not of the actual reverse transition. The advertised first-principles result ('CNNs can approximate reverse diffusion transitions with arbitrarily low error') reduces to a definition that discards the dependence on the random draws, so the guarantee is an artifact of the construction rather than a derived theorem.

full rationale

This paper contains no self-citation chain: the diffusion construction [1, 17], the universal approximation theorems [32, 33], and the uncertainty-weighting [34] all cite external work, so the self-citation, uniqueness-imported-from-authors, and ansatz-smuggled-via-citation patterns do not apply. The circularity is internal to the advertised theoretical guarantees. First, Lemma 1's proof that CNN reverse diffusion 'enables mode coverage' assumes exactly the property to be established when it states that 'fθ maps each xt toward its corresponding x, which comes from a particular mode,' and then reads off the support-inclusion conclusion; the conclusion is also forced by construction, since the reconstruction objective samples x from all K modes (π_k > 0), so any successful reconstruction has outputs ranging over the union of mode supports regardless of the CNN structure. Second, Lemma 2's 'reverse mapping' g(xt) is manufactured by algebraically solving the forward diffusion equation, but the resulting expression still contains the random draws x0_j and εt, which are not determined by xt; no function g: R^M → R^M exists, so the claimed Lipschitz constant and the sup-norm bound are properties of a constructed artifact rather than a first-principles derivation. The headline empirical claim (CDNet improves SOTA deep classifiers on the UCR Archive and simulated data) is benchmarked against external baselines and is therefore not circular; it is, however, reported inconsistently (e.g., ECG200 CDNet accuracies differ between Tables 1 and 2, and several δ columns in Table 1 do not equal the CDNet-minus-baseline differences), which is a correctness risk rather than a circularity. The paper's own limitation statement — 'we agree explicit comparisons to contrastive-augmentation baselines would strengthen our contributions' — is weighed here as an acknowledged gap in establishing the augmentation mechanism's independent value, but it does not itself constitute circularity. Overall, the empirical core retains independent content, while the theoretical 'guarantees' — a headline contribution — reduce by construction, yielding partial circularity (score 6).

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The method depends on the reverse CNNs being able to denoise interpolated mixtures, so the key premises are the compact-Sobolev domain assumption, the CNN approximation theorem, and the false determinism premise in Lemma 2. Lemma 1 adds an unproved convergence-to-mode-preservation step. None of the hyperparameters needed to reproduce the method are reported.

free parameters (4)
  • diffusion step count T and noise schedule β_t
    Reported as existing but values are never given; the reverse CNNs f^t_θt are indexed by t, so T defines the number of networks and the noise level per step.
  • forward noise scale σ
    ε ~ N(0, σ^2I) in the forward process; controls the denoising difficulty and generated-sample diversity; no value reported.
  • triplet margin α and SNN temperature τ
    Hand-set constants in LTriplet and LSNN; no values reported, and they directly control the contrastive training signal.
  • uncertainty weights σ_CE, σ_SNN, σ_Triplet
    Learnable loss weights in the composite loss; fitted during pre-training but final values are not reported.
assumptions (5)
  • domain assumption Clean time series belong to a compact subset of the Sobolev space W^{1,2}([0,1]) with bounded first derivatives.
    Invoked in Lemma 2 to make discretized time series compact so the CNN universal approximation result applies; excludes discontinuous signals common in real time series.
  • standard math 1D CNNs can universally approximate Lipschitz functions on compact Sobolev-type domains (Yarotsky 2017, Zhou 2020).
    This is the cited theorem [32, 33]; it applies only if the target function is well-defined and Lipschitz, which Lemma 2 fails to establish because the inverse is not a function of xt.
  • ad hoc to paper The reverse transition xt-1 is a deterministic Lipschitz function of xt with constant 1/sqrt(1-βt).
    This is the fatal premise: the derived expression contains x0_j and εt, which are not determined by xt; the premise is false, so Lemma 2 collapses.
  • ad hoc to paper SGD converges to a stationary point of the reconstruction loss and that point induces the claimed mode-coverage support property.
    Lemma 1's proof assumes convergence and then asserts fθ maps each xt to 'its corresponding x'; the MMSE target for interpolated mixtures is a blend, not a mode-specific point, so the support conclusion is not established.
  • domain assumption Linear interpolation between two same-class samples plus Gaussian noise stays on or near the class manifold.
    The positive and negative sample generation assumes interpolated mixtures remain semantically meaningful (e.g., phase-shifted heartbeats); this is not generally true and is untested.
invented entities (1)
  • Learned instance-to-instance reverse diffusion transitions (positive/negative contrastive trajectories)
    purpose: Generate extra training samples for the classifier by undoing within-class and across-class interpolations
    These constructs are internal to CDNet; the only external handle is the accuracy tables, which are internally inconsistent and are not a falsifiable prediction independent of the method.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Contrastive Diffusion-based Network (CDNet) for Time Series Classification." pith.science (2026). https://pith.science/paper/2JMXWUP7

@misc{pith2026250721357,
  author       = {Pith},
  title        = {Pith review of: A Contrastive Diffusion-based Network (CDNet) for Time Series Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2JMXWUP7}},
  note         = {Machine review of arXiv:2507.21357}
}
read the original abstract

Deep learning models are widely used for time series classification (TSC) due to their scalability and efficiency. However, their performance degrades under challenging data conditions such as class similarity, multimodal distributions, and noise. To address these limitations, we propose CDNet, a Contrastive Diffusion-based Network that enhances existing classifiers by generating informative positive and negative samples via a learned diffusion process. Unlike traditional diffusion models that denoise individual samples, CDNet learns transitions between samples--both within and across classes--through convolutional approximations of reverse diffusion steps. We introduce a theoretically grounded CNN-based mechanism to enable both denoising and mode coverage, and incorporate an uncertainty-weighted composite loss for robust training. Extensive experiments on the UCR Archive and simulated datasets demonstrate that CDNet significantly improves state-of-the-art (SOTA) deep learning classifiers, particularly under noisy, similar, and multimodal conditions.

Figures

Figures reproduced from arXiv: 2507.21357 by the authors.

Figure 1
Figure 1. The ECG200 dataset from the UCR Archive. The top row and bottom row correspond to two classes, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An overview of CDNet. As shown in [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. A sample diffusion process between two randomly selected samples within a class. The top row is the forward [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Critical Difference (CD) diagram comparing CDNet variants on top of InceptionTime across all UCR Archive [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: CD diagram comparing CDNet variants on top of 1DCNN across all binary UCR Archive datasets. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: CD diagram of CDNet variants with LSTM_FCN on the UCR Archive. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: CD diagram of CDNet on top of InceptionTime and SOTA models on the UCR Archive. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: CD diagram of CDNet on top of 1DCNN and SOTA models on the UCR Archive. [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: CD diagram of CDNet on top of LSTM_FCN and SOTA models on the UCR Archive. [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 10
Figure 10. Figure 10: Average accuracy of BaseCNN and CDNet_CNN and improvement wrt noise level. A higher noise level [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]
Figure 11
Figure 11. Figure 11: Average accuracy of BaseCNN and CDNet_CNN and improvement wrt class similarity level. A higher [PITH_FULL_IMAGE:figures/full_fig_p010_11.png]
Figure 12
Figure 12. Figure 12: Average accuracy of BaseCNN and CDNet_CNN and improvement wrt class multimodal level. A higher [PITH_FULL_IMAGE:figures/full_fig_p011_12.png]
Figure 13
Figure 13. Figure 13: A sample simulation plot. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_13.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 39 canonical work pages

  1. [1]

    Card: Classification and regression diffusion models

    Xizewen Han, Huangjie Zheng, and Mingyuan Zhou. Card: Classification and regression diffusion models. Advances in Neural Information Processing Systems, 35:18100–18115, 2022

    work page 2022

  2. [2]

    shapeDTW: Shape dynamic time warping

    Jiaping Zhao and Laurent Itti. shapeDTW: Shape dynamic time warping. Pattern Recognition, 74:171–184, 2018

    work page 2018

  3. [3]

    catch22: CAnonical Time-series CHaracteristics

    Carl H Lubba, Sarab S Sethi, Philip Knaute, Simon R Schultz, Ben D Fulcher, and Nick S Jones. catch22: CAnonical Time-series CHaracteristics. Data Mining and Knowledge Discovery, 33(6):1821–1852, 2019

    work page 2019

  4. [4]

    Scalable dictionary classifiers for time series classification

    Matthew Middlehurst, William Vickers, and Anthony Bagnall. Scalable dictionary classifiers for time series classification. In IDEAL, pages 11–19. Springer, 2019

    work page 2019

  5. [5]

    Fast and accurate time series classification with weasel

    Patrick Schäfer and Ulf Leser. Fast and accurate time series classification with weasel. In CIKM, pages 637–646. ACM, 2017

    work page 2017

  6. [6]

    HIVE-COTE 2.0: A new meta ensemble for time series classification

    Matthew Middlehurst et al. HIVE-COTE 2.0: A new meta ensemble for time series classification. Machine Learning, 110(11):3211–3243, 2021

    work page 2021

  7. [7]

    MultiRocket: Multiple pooling operators and transformations for fast and effective time series classification

    Chang Wei Tan et al. MultiRocket: Multiple pooling operators and transformations for fast and effective time series classification. Data Mining and Knowledge Discovery, 36(5):1623–1646, 2022

    work page 2022

  8. [8]

    Fast and accurate time series classification through supervised interval search

    Nestor Cabello et al. Fast and accurate time series classification through supervised interval search. In IEEE ICDM, pages 948–953, 2020

    work page 2020

Show all 40 references
  1. [9]

    The freshprince: A simple transformation based pipeline time series classifier

    Matthew Middlehurst and Anthony Bagnall. The freshprince: A simple transformation based pipeline time series classifier. In ICPRAI, pages 150–161. Springer, 2022

    2022

  2. [10]

    SFA: A symbolic fourier approximation and index

    Patrick Schäfer and Mikael Högqvist. SFA: A symbolic fourier approximation and index. In EDBT, pages 516–527, 2012

    2012

  3. [11]

    Time series feature extraction using scalable hypothesis tests.Neurocomputing, 307:72–77, 2018

    Maximilian Christ et al. Time series feature extraction using scalable hypothesis tests.Neurocomputing, 307:72–77, 2018

    2018

  4. [12]

    Self-supervised learning for semi-supervised time series classification

    Shayan Jawed, Josif Grabocka, and Lars Schmidt-Thieme. Self-supervised learning for semi-supervised time series classification. In PAKDD, pages 499–511. Springer, 2020

    2020

  5. [13]

    Self-supervised contrastive representation learning for semi-supervised time-series classification

    Emadeldeen Eldele et al. Self-supervised contrastive representation learning for semi-supervised time-series classification. IEEE TPAMI, 2023

    2023

  6. [14]

    Barlow twins: Self-supervised learning via redundancy reduction

    Jure Zbontar et al. Barlow twins: Self-supervised learning via redundancy reduction. InICML, pages 12310–12320, 2021

    2021

  7. [15]

    Bootstrap your own latent: A new approach to self-supervised learning

    Jean-Bastien Grill et al. Bootstrap your own latent: A new approach to self-supervised learning. NeurIPS, 33:21271–21284, 2020

    2020

  8. [16]

    The UCR time series archive

    Hoang Anh Dau et al. The UCR time series archive. IEEE/CAA JAS, 6(6):1293–1305, 2019

    2019

  9. [17]

    Diffusion models in vision: A survey

    Florinel-Alin Croitoru et al. Diffusion models in vision: A survey. IEEE TPAMI, 45(9):10850–10869, 2023

    2023

  10. [18]

    InceptionTime: Finding AlexNet for time series classification

    Hassan Ismail Fawaz et al. InceptionTime: Finding AlexNet for time series classification. DMKD, 34(6):1936– 1962, 2020

    1936

  11. [19]

    Convolutional neural networks for time series classification

    Bendong Zhao et al. Convolutional neural networks for time series classification. Journal of Systems Engineering and Electronics, 28(1):162–169, 2017

    2017

  12. [20]

    Multivariate LSTM-FCNs for time series classification

    Fazle Karim et al. Multivariate LSTM-FCNs for time series classification. Neural Networks, 116:237–245, 2019

    2019

  13. [21]

    Deep learning for time series classification: a review

    Hassan Ismail Fawaz et al. Deep learning for time series classification: a review. Data Mining and Knowledge Discovery, 33(4):917–963, 2019

    2019

  14. [22]

    Scalable diverse model selection for accessible transfer learning

    Daniel Bolya, Rohit Mittapalli, and Judy Hoffman. Scalable diverse model selection for accessible transfer learning. NeurIPS, 34:19301–19312, 2021

    2021

  15. [23]

    Time-series forecasting with deep learning: a survey

    Bryan Lim and Stefan Zohren. Time-series forecasting with deep learning: a survey. Philosophical Transactions of the Royal Society A, 379(2194):20200209, 2021

    2021

  16. [24]

    Machine learning advances for time series forecasting

    Ricardo P Masini, Marcelo C Medeiros, and Eduardo F Mendes. Machine learning advances for time series forecasting. Journal of Economic Surveys, 37(1):76–111, 2023

    2023

  17. [25]

    An examination of the state-of-the-art for multivariate time series classification

    Bhaskar Dhariyal et al. An examination of the state-of-the-art for multivariate time series classification. In ICDMW, pages 243–250. IEEE, 2020

    2020

  18. [26]

    Using dynamic time warping to find patterns in time series

    Donald J Berndt and James Clifford. Using dynamic time warping to find patterns in time series. In KDD, pages 359–370, 1994. 18 A PREPRINT - S EPTEMBER 13, 2025

    1994

  19. [27]

    CNN approaches for time series classification

    Lamyaa Sadouk. CNN approaches for time series classification. Time Series Analysis—Data, Methods, and Applications, volume 5, pages 57–78, 2019

    2019

  20. [28]

    Scalable classification of univariate and multi- variate time series

    Saeed Karimi-Bidhendi, Faramarz Munshi, and Ashfaq Munshi. Scalable classification of univariate and multi- variate time series. In IEEE Big Data, pages 1598–1605, 2018

    2018

  21. [29]

    Diffusion models: A comprehensive survey of methods and applications

    Ling Yang et al. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4):1–39, 2023

    2023

  22. [30]

    A survey on deep semi-supervised learning

    Xiangli Yang et al. A survey on deep semi-supervised learning. IEEE TKDE, 35(9):8934–8954, 2022

    2022

  23. [31]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840–6851, 2020

    2020

  24. [32]

    Error bounds for approximations with deep ReLU networks

    Dmitry Yarotsky. Error bounds for approximations with deep ReLU networks. Neural Networks, 94:103–114, 2017

    2017

  25. [33]

    Universality of deep CNNs for function approximation on Sobolev spaces

    Chaoyue Zhou. Universality of deep CNNs for function approximation on Sobolev spaces. arXiv preprint arXiv:2002.11520, 2020

    2002 arXiv

  26. [34]

    Multi-task learning using uncertainty to weigh losses

    Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses. In CVPR, pages 7482–7491, 2018

    2018

  27. [36]

    Finding order in chaos: A novel data augmentation method for time series in contrastive learning

    Berken Utku Demirel and Christian Holz. Finding order in chaos: A novel data augmentation method for time series in contrastive learning. In NeurIPS 2023 Poster, 2023

    2023

  28. [37]

    Time-series representation learning via temporal and contextual contrasting

    Emadeldeen Eldele et al. Time-series representation learning via temporal and contextual contrasting. arXiv preprint arXiv:2106.14112, 2021

    2021 arXiv

  29. [38]

    Time-series representation learning via temporal and contextual contrasting

    Emadeldeen Eldele et al. Time-series representation learning via temporal and contextual contrasting. In IJCAI, pages 2352–2359, 2021

    2021

  30. [39]

    TS-TCC: Time-series representation learning via temporal and contextual contrasting

    Burak Demirel, Jungtaek Baek, and Sungroh Choi. TS-TCC: Time-series representation learning via temporal and contextual contrasting. In NeurIPS 2023, 2023

    2023

  31. [40]

    TS2Vec: Towards universal representation of time series

    Zhihan Yue et al. TS2Vec: Towards universal representation of time series. In AAAI, pages 8980–8987, 2022

    2022

  32. [41]

    Augmentations for self-supervised time-series representation learning

    Zhaohui Wang et al. Augmentations for self-supervised time-series representation learning. In ICLR, 2021. 19

    2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.